Re: Accuracy of Mathematical Functions

On Mon, Sep 25, 2023 at 08:20:25PM +0200, Dimitry Andric wrote:
> On 25 Sep 2023, at 15:50, Paul Zimmermann <Paul.Zimmermann@inria.fr> wrote:
> > 
> > I hope this is not off-topic for this list (Technical discussions relating
> > to FreeBSD).
> 
> The freebsd-numerics@ list might have been a better match, but it
> receives very low traffic, and the audience on this list will be larger.

I suggested Paul post here as numerics@ has had 32 post since
2021 and many of those are due to someone assigning a bug
report to numerics.  


> > We have updated our comparison:
> > 
> > https://members.loria.fr/PZimmermann/papers/accuracy.pdf
> > 
> > This new update includes for the first time the FreeBSD math library,
> > whose accuracy is quite good, except:
> > 
> > * single precision: the Bessel functions, lgammaf, cospif, sinpif, tanpif, powf
> > * double precision: the Bessel functions, lgammaf, tgammaf, cospi, sinpi,
>>  tanpi, pow

lgammaf and tgammaf are single precision, but it is certainly possible
that there are issues.  I'll take a look when I have time.

>> * double-extended precision: erfcl, lgammal, tgammal, cospil, sinpil, tanpil,
>>  powl
>> 
>> Some issues have already been fixed in the development version by Steve
>> Kargl (we used FreeBSD 13.2).
> 
> Very interesting paper! Of course we are always interested in
> improvements for libm, and Steve semi-regularly posts patches in our bug
> tracker. (Steve's no longer a committer, but usually these get committed
> by others quickly enough.)

The issues with half-cycle trig function (i.e., cospi and friends)
should be fixed with git 2d3b0a687 on the main branch.  It looks
like kib merged the patch into stable/13 with git 6fe5d4d8.

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=272742

-- 
Steve